Zero-byte disk writes. Zero gradient training epochs. QTensor compresses foundation-class LLMs to sub-3GB edge silicon and consumer GPUs using deterministic linear algebra and subspace residual bridges.
Instantaneous closed-form linear algebra compression executed with zero gradient training epochs.
Methodology: Evaluated live across official benchmark suites: HumanEval+ (54.88% Pass@1 on Qwen2.5-Coder-7B), LiveCodeBench v6 (175 problems, 81.1% AST, 58.1% Easy Pass@1 on Gemma 4 E4B-IT, 30.0% Pass@1 on Gemma 2 9B), MMLU 5-shot (72.06% Macro-Average / 95.82% retention on Gemma 2 9B QTensor), and Llama 3.2 3B (0.80 exit cosine similarity). Full audited telemetry maps and Triton kernel specifications are available in the Technical Brief.
Hardware Verification
Production Deployment Matrix
Real measured footprints, benchmark accuracy, and target silicon across verified QTensor compilations.
Model Architecture
Base Footprint
QTensor Footprint
Reduction
Verified Benchmark
Target Hardware
Access
Llama 3.2 3B QTensor v2
Protected Semantic Corridor Compilation
6.00 GiB (BF16)
2.98 GiB (v1) / 5.60 GiB (v2)
-50.4% / -6.7%
0.8008 Cosine Sim
Zero Repetition Loops
ARM Cortex / Apple Silicon / 4GB Edge
Qwen 2.5 Coder 7B QTensor
Tier 1 Analytical Corridor Compilation
14.2 GB (FP16)
3.02 GB
-78.7%
54.88% Pass@1
HumanEval+ Sandbox
4GB Industrial Edge / Jetson Orin
Gemma 4 E4B-IT QTensor
Analytical Subspace Bridge Compilation
16.0 GB (>16GB OOM)
15.15 GB Peak VRAM
Fits 16GB VRAM
81.1% AST Correct
58.1% Pass@1 (LiveCodeBench v6)
RTX 5080 (16GB) / Consumer GPUs
Gemma 2 9B QTensor
Closed-Form Analytical Subspace Decomposition
18.5 GB (BF16)
5.71 GB static
11.10 GB peak VRAM
-69.1% static
72.06% MMLU (95.82% ret.)
30.0% Pass@1 (100% AST LiveCodeBench)
RTX 5080 (16GB) · 14.7 TPS
0.0% WDDM PCIe paging
* All models benchmarked zero-shot without fine-tuning or quantization calibration sets.
View Live Telemetry ↓
Preserves critical semantic embedding spaces and logit projection boundaries. QTensor isolates rank optimization away from sensitive boundary layers while maintaining 100% native attention mechanics across the network.
Topological Isolation · Zero Manifold Distortion
Linear Algebra
02
Subspace Residual Bridging
Standard low-rank factorization causes severe activation collapse and repetitive looping. QTensor employs proprietary analytical residual bridging to dynamically compensate for non-linear energy loss, eliminating repetition loops on long-context code generation.
Topological buffer expansion with analytical corridor compilation. Full AST compliance and 0% repetition loops.
GATED3.02 GB
qwen-2.5-coder-7b-qtensor
Sub-4GB edge compilation retaining 54.88% Pass@1 on HumanEval+. Zero gradient training.
GATED5.71 GB (11.10 GB Peak)
gemma-2-9b-it-qtensor
Closed-form analytical subspace decomposition. 72.06% MMLU 5-shot retention (95.82% vs. BF16) and 30.0% Pass@1 on LiveCodeBench v6 with 14.7 TPS on RTX 5080.
GATED4.85 GB
QTensor-Llama-3.1-8B
Enterprise edge compilation for autonomous and embedded systems. High-entropy channel calibration.
QTensor's core compiler is supported by ZeroFold's engineering practice. We deliver custom kernel optimization, edge hardware integration, and sovereign production pipelines for enterprise, defense, robotics, and embedded systems.
* Evaluation weights and proprietary Triton kernel sources are released under Enterprise MNDA.
Technical Brief Access
Request Evaluation License
Access the complete QTensor v2.0 evaluation license, technical specifications, and gated model weights.
✓
Request Received
Your evaluation license request has been securely submitted to consult@qtensor.com.au. An engineering representative will review your target hardware specifications and contact you shortly.